# Researchers report 8B agent model result with EvoHarness-RL

_Friday, August 28, 2026 at 12:33 PM EDT · AI, Science · Latest · Tier 2 — Notable_

![Researchers report 8B agent model result with EvoHarness-RL — Primary](https://images.ctfassets.net/jdtwqhzvc2n1/74qlOTYyXbU8te5ugdb5DX/f6696a755ed1db9f3a43cdb5ec9219c8/AI_harness.jpg?w=800&q=75)

Meta AI and University of Illinois Urbana-Champaign researchers report that their EvoHarness-RL training framework took a Qwen3-8B model to a 96.9% average success rate on the ALFWorld benchmark. The reported score exceeded the article's cited Claude Opus 4.5 result of 96.4% and was 49.0 points above the ReAct baseline. The evaluation concerns a text-based multi-step task, not a production deployment.

## Sources

- [VentureBeat](https://venturebeat.com/orchestration/meta-researchers-taught-an-8b-ai-model-to-match-claude-opus-4-5-without-the-frontier-price-tag)

---
Canonical: https://techandbusiness.org/newswire/PW5d8t6hPyQZ22qccVs_AU
Retrieved: 2026-08-28T18:59:35.231Z
Publisher: Tech & Business (techandbusiness.org)